Large language models are typically trained with next-token prediction (NTP), which rewards predicting specific tokens rather than explicitly accounting for …
机构:Stanford
来源:arXiv 2610.29843 | AI4Papers 论文推荐平台